After the Moat

What can we do that machines can’t do? And what value do we bring to a world where AI systems are better than us at a wide range of tasks?

We talk a lot these days, in journalism and beyond, about “defensible moats” — from business models to human skills — that will keep us employed as AI reshapes how news and information gets produced, distributed and consumed. But the technology is moving so fast that many of those moats are drying up quite quickly.

So perhaps we should reframe our analogy. Instead of playing defense — trying to identify human-only skills impervious to AI — we should instead think about what we need to learn in a world where it’s possible that AI becomes better (or at least more efficient) than us at doing almost everything we currently do? What do we need to master so that we can exercise judgement and control over these systems long term, to ensure they work in the service of our goals and our values?

This isn’t a theoretical exercise for me. Outside of my day job at the Tow-Knight Center at CUNY, I also teach investigative reporting at the (ahem!) other New York City graduate journalism school that starts with C. Like many professors, this year I’ve been re-examining what and how to teach emerging journalists as AI changes the processes of reporting (and of learning).

My summer class started this week, and in planning for it I found myself scrapping the lesson plan I’ve used on backgrounding people and companies for half a decade — a skillset I’ve long considered an important part of investigative journalism — after I built an AI-powered workflow for making a corporate backgrounding dossier in under an hour.

The dossier workflow shrank hours, if not days, of work to a few minutes, and the more layers of database queries (through API and MCP calls) that I added to it, and the better I made the final presentation (with footnotes, timelines, and follow-up leads), the more it forced me to examine what my goals truly are when I teach students how to background.

The way I learned to use public records to background, and the way I’ve taught it for years, means combing through corporate filings, nonprofit 990s, court records, permits, contracts, fines and more to follow the money, map networks and relationships, and build timelines that help prove when and how wrongdoing happened. That’s a dull, laborious process, but it’s useful beyond just the information it uncovers.

It builds muscle memory and pattern recognition, an institutional fluency born from the drudgery of looking at records. It’s a specialized skill that historically often set investigative reporters apart — let’s call it the “records moat” — knowing where public records are kept, how to get at them and pull information out, what’s relevant or unusual, how one record might connect to another to provide a lead or proof, and how to organize the information to show a larger picture with an auditable paper trail.

This summer, I wanted to see how AI could fit into that lesson. A demo of Enigma’s Gov Archive MCP server at Media Party NYC last month got me started. (An MCP — Model Context Protocol — server lets an AI assistant plug directly into a data source, using plain language rather than code. This one pulls records from a collection of thousands of U.S. government databases.) I looked up a local ice cream chain that had been through a bankruptcy and a few ownership changes. It found dozens of database results, from corporate acquisitions and storefront openings and closings across state lines, to loans, licenses, overdue fire-code fines, and more. It only took a few minutes to use AI to turn the results into a footnoted dossier, complete with a corporate timeline, key individuals, an organizational chart, with reporting leads and data gaps at the top.

But then I began adding to it, layering in API searches of EDGAR, OpenCorporates, RECAP, ProPublica’s Nonprofit Explorer, the Wayback Machine and a half dozen other data sources I regularly search when backgrounding (the ice cream chain did not appear in all of them). The whole workflow runs in minutes now, almost all the searches for free, with the few paid ones just a few cents per search.

The time savings were the obvious part. But the aha moment was figuring out how to format the outputs — by creating a template of how I wanted the results of the data queries filtered and presented based on what I know is relevant to my work. I’ve made dozens, if not hundreds, of these manually, but to make a template, I had to spell out what mattered to me, encoding a significant portion of my news judgement into the AI process.

I was teaching the workflow the kinds of information I wanted it to surface, what to leave in underlying source files, and the types of connections and suggested next steps I find valuable. I built in places for me to assess and trace the AI outputs, but I also reduced the number of judgement calls I need to make for each individual company I background.

In doing so, I didn’t remove human judgement from the dossier creation process; I moved it upstream. This is happening across AI workflows of all kinds — when you establish a repeatable process, the judgement calls made in the planning of that process often become far more important than the ones you make in the execution.

But moving judgement upstream creates a new problem. If I hand this dossier generator to my students, how do they ever develop that judgement? And do they even need it in this new world?

I think they do, but not because it can’t be replaced.

Let’s assume the work can be replaced, all of it. All it takes is one human, one time creating a useful process. Assume the dossier, the leads, even the connections between them come back as sharp as anything I’d have built by hand. At that point, what’s the value of a human doing the task if a machine can do it more efficiently?

So if that backgrounding skill is no longer a moat, what is? What do my students need to master?

They need to learn reasoning processes, critical thinking and how to exercise judgement when they use AI, and to carry what they learn into the moments that don’t involve technology at all. Records work helps reporters get a feel for the shape of what they’re investigating — to understand complex systems and the players and pressure points within them. The value of the manual task is much more than the actual facts returned, but also the interpretation and judgement learned through having to sift, filter and absorb information at the speed of human reading.

Plenty of the journalistic process still depends on the comprehension that records work used to build. Reporters still have to connect dots in an interview, recognize leads, ask follow-ups on the fly, weigh newsworthiness, and defend their sourcing and accuracy.

So I’m not handing my students the workflow — instead the lesson will focus on helping them understand the records landscape and how to create their own workflow for it. I’ve shifted from building a records moat to equipping them with a records map and navigation guidelines. They’ll have to do some searching by hand, but they’ll also have to come up with their own parameters for making a dossier template from data queries.

Both are important, because together they provide the calibration for the parts humans can’t, or won’t, delegate.

The goalposts have moved. It’s no longer about teaching students to research; it’s about teaching them to oversee, interpret and assess research workflows and outputs — to trace where the information came from, interrogate the judgements baked into the results and build the reflexes to always ask what might be missing and where to look next. Because somewhere in there, someone always makes a call: a reporter deciding what they think is useful, an MCP designer deciding what surfaces, an engineer at a frontier model company deciding the defaults for relevance and credibility for everyone.

The hard skills and the workflows will keep changing — they’ve changed enormously in just six months. The thing I’m confident in is that, no matter how good these systems get at executing judgements against a set of parameters, the human ability to make those judgements in the first place still matters.

It’s not a question of whether the machine can do it better than us; it’s a question of what cognitive skills are important for humans regardless of our environment and tools. I’m betting on judgement. In a few weeks, we’ll see how the new lesson goes.